[Dynamite] (no subject)

[Ian Holmes]

On Wed, 26 Jul 2000, Ewan Birney wrote:

> 
> > ...this is not total fluff---the point is that loop states change the
> > actual functionality of the model, whereas singlet states that aren't loop
> > states (like INSERT_QUERY_* above) don't change the functionality (the
> > scores returned by algorithms will be the same; it will just shave a bit
> > off the running time of the algorithm because you don't store a whole
> > bunch of empty cells in the matrix).
> 
> 
> No. Because there are singlet->singlet potential transitions and the
> viterbi or posterior score when we jump into the polymer matrix is data
> dependent and path (through singlet) dependent.
> 

Would you mind illustrating this with an example please? I am not
convinced that singlet states can't be auto-detected and I think I may not
be understanding your point.

Either way, I have now come round to the idea that we should label all
special states, pretty much exactly the way Dynamite does, except that we
should call them "singlet" not "special". And yes, on reflection "singlet"
is a much better name than "loop".

So, just to be clear, a special state looks like this:

	<state name="fred"> <singlet chain="target"/> </state>

whereas a regular state looks like this

	<state name="jim"> </state>

or, equivalently

	<state name="jim"/>

For now, I think we should also (being nice to Guy) stipulate that the
state order that passes through the XML layer (from me to Guy) can be used
unchanged, i.e. I will pre-sort the states so that singlets and null
transitions are handled nicely.

Ian